Venture Guides Dinner Series: Solving the Barriers to Enterprise AI Adoption

75% of Enterprises experimented with generative AI in 2023, but only 9% widely adopted the technology. Given Deep Learning has transformed how we think about computing capabilities, we went to the experts to understand what's holding businesses back.  

In our first official Venture Guides Dinner series event, we wanted to explore the dynamics around enterprise AI adoption, known barriers, and novel solutions to overcome them.  So we brought in experts including executives from Oracle, Capital One, and Google; MIT and Microsoft Researchers; and Founders of companies including Liquid AI, AppMap, and ElastiFlow for dinner, discussion, and collective learning.

Let's dive into our three key takeaways:  

1. Enterprises seek AI applications that achieve outcomes, not just outputs 

When it comes to deep learning, much of the excitement is driven by Generative AI. ChatGPT, Midjourney, and DALL-E were excellent at creating outputs (essays, images, etc.), but they didn’t necessarily accomplish end-to-end business workflows or outcomes.  

That’s the shift that’s needed for enterprises. While they do experiment with today’s “output-focused” AI tools for productivity gains (writing marketing copy, boilerplate code fixes, etc.), the treasure chest lies in outcomes.  

When AI can complete business-centric workflows in a comprehensive, trustworthy manner (understanding and processing legal contracts, launching full features into production, etc.), we’ll see a seismic shift where enterprises move from experimentation to major adoption.  

But how do we get there? 

2. Agentic automation could be the next step in AI development 

An AI Agent (or Intelligent Agent) is an AI-driven model that achieves tasks autonomously. It senses its environment, makes decisions, and takes actions by itself to achieve specific objectives.  

Business workflows are the atomic units of enterprises. Agents could invigorate enterprise adoption by tackling and holistically optimizing these very atomic units. Imagine setting a marketing campaign goal, setting an AI Agent on it, and letting it continuously optimize itself as it learns what copy/visuals work and don’t!  

AI Agents are built by enhancing LLMs with more context, data, and connectors. They essentially take the “brains” of Deep Learning, then equip it with the capabilities to do end-to-end tasks.  

But Agents still tie back to something expensive and opaque: Neural Networks. As a result, we still run into a familiar barrier to enterprise adoption: cost.  

3. Deep Learning models are compute intensive and it’s near impossible to precisely explain how they generate outputs. But novel methodologies and neural network architectures could unlock adoption. 

Alexander Amini, co-founder and Chief Science Officer of Liquid AI, and a AI Researcher at Massachusetts Institute of Technology (MIT) shares how his team has rearchitected neural networks.

The high costs of AI (both when training models and inferencing them) and its lack of explainability both trace back to a common source: today’s Neural Networks are enormous (GPT-4 is rumored to have 1.74 trillion parameters), making them incredibly expensive to run and near-impossible to explain (in a setup of 1.74T predictors, how would we tease out what’s generating what output?). 

As the Liquid AI team shared with us at the dinner, they've recently rearchitected neural networks: they’re calling this evolution Liquid Neural Networks. Liquid Neural Networks enhance each parameter to make neural networks faster, more adaptable, and smaller. In fact, they were able to use just 19 nodes (as opposed to 100,000) in an autonomous vehicle experiment.  

Meanwhile, we’re seeing leaps in research at MIT’s Han Lab that make existing Neural Networks more efficient (for example, TinyML), explainable, and uncertainty-aware. Between experimenting with novel neural network architectures and identifying novel techniques in efficient machine learning, founders and researchers are aiming to quickly demolish the cost, trust, and efficiency barriers to enterprise adoption. 

Ultimately, we know AI will infiltrate the enterprise. The question is just when and how.  

If you’re building in the space, operating in it, or have strong opinions around AI adoption in the enterprise, we’d love to hear from you and have you join our next event!